Contact Force Control
David Johnston,
yPing Zhang,
zJohn Hollerbach, and Stephen Jacobsen
Center for Engineering Design and
zDept. Computer Science,
Univ. Utah, Salt Lake City, UT 84112
y
Center for Intelligent Machines, McGill Univ., Montreal, QC H3A 2A7
Abstract
A full tactile sensing suite for the nger segments and palm of the Utah/MIT Dextrous Hand is pre-sented. The rubber-based sensors employ capaci-tance sensing and oating electrodes in the top layer, and contain local electronics for excitation, ltering, analog-to-digital conversion, and serial communica-tion. Experimental results on static, dynamic, and spatial properties are presented. Use of the tactile sen-sor in contact force control is demonstrated.
1 Introduction
Tactile sensor development remains an enigma for robotics. While many laboratory prototypes have been proposed [2, 4, 5, 9], and occasionally some com-mercial product appears, the perception nevertheless is that there are not good o-the-shelf sensors to be found. Consequently multi-ngered robot hands or other grippers mostly do without tactile sensors, while those that do incorporate them are not entirely satis-factory.
One problem is that there is a huge step from labo-ratory prototype to commercial product. Robustness, reliability, complexity, manufacturability and cost can be insurmountable obstacles for what otherwise looks good in the laboratory. Another problem is that the utility of tactile sensing in robotics has not been deni-tively demonstrated. It is an article of faith among many robot investigators that tactile sensing is impor-tant, but there is the question of having good tactile sensors available to prove it.
This paper presents a full tactile sensing system for the Utah/MIT Dextrous Hand, which is now commer-cially available through Sarcos Inc. (Salt Lake City, UT). The tactile sensing system covers all nger seg-ments including curved ngertips and the palm (Fig-ure 1). The tactile sensors slip on to the nger seg-ments, and while developed initially for the UMDH could readily be recast for other nger segment shapes.
Figure 1: Tactile sensor suite for all nger segments and palm of the Utah/MIT Dextrous Hand.
The next section sketches the components and con-struction of the sensors. Thereafter experimental re-sults for a single tactile pad are presented that charac-terize static, dynamic, and spatial properties. Finally, the use of a tactile sensor in contact force control is demonstrated.
2 System Description
The sensor technology is comprised of capacitance sensing and arrays, dened in rubber layers. The de-velopment is similar to that proposed by [1], in that oating electrodes (Figure 2) are employed in the top layer rather than conductive strips as in [3, 11, 10]. The reason is for greater robustness, because of the possibility of contact breakage with conductive strips. The drawback is a smaller signal. Initial designs for this tactile sensor were presented in [7, 8]. This pa-per presents expa-perimental results with the nal proto-types.
The tactile system mounted on the forenger and palm of the UMDH is shown in Figure 3. Individual tactile elements (tactels) are on 2.77mm centers; this distance was dictated by size requirements to achieve suciently high output. The at array in the palm
SENSE EXCITATION SENSE EXCITATION FLOATING CONDUCTOR (a) (b)
Figure 2: Capacitor tactile construction by (A) con-ductive strips in top layer or (B) oating electrodes in top layer.
Figure 3: Tactile sensors attached to index nger and palm
contains 64 sensors, the nger link 1 (proximal) array contains 76 sensors, the nger link 2 array contains 36 sensors, and the ngertip array contains 58 sensors. The ngertip array is the more complex to fabricate, because of the curved ngertip.
Figure 4 shows the components of the tactile sens-ing arrays for one nger. Shown are molds for the n-ger segment shapes, PC boards for local electronics, connector strips, coplanar plate arrays, top oating electrodes, and covers. The covers serve to insulate against stray capacitance, and also contain dimples for stress concentration. Tactile array assemblies are attached to nger segments via screws through a back plate.
Electronics to drive each sensor array are located very close to each tactile pad in the nger segments.
Figure 4: Tactile sensor components for one nger. These electronics include excitation electronics, ana-log lter, Motorola 6811 microcontroller, and inter-face electronics. The interinter-face electronics communi-cate over a serial line to an intermediate electronic interface mounted on the back of the hand, and then to an interface card on a VME bus (Figure 5). Each of 13 arrays transmits data at an eective rate of 17,000 tactels per second. The bandwidth of each array there-fore depends on the number of sensors in that array (Table 1). CONTROLLER & INTERFACE CARD VME BUS DATA I/O GND CLK +5v -5v
RANDOM ACCESS ADDRESSING:
• 13 ARRAYS
SEQUENTIAL ADDRESSING:
• 58 Sensors / Fingertip • 36 Sensors / Link 2 • 76 Sensors / Link 1 (proximal) • 64 Sensors / Palm
744 TOTAL SENSORS WORKSTATION
DEXTROUS® HAND LINKS
Figure 5: System conguration.
3 Experimental Results
3.1 Experimental Setup
The experimental setup is shown in Figure 6. A at palm array was used for testing. A Bruel and Kjaer voice coil motor exerts force on a tactel through a probe. The tactile pad is located under the probe by an x-y stage. A combined LVDT-LVT measures po-sition and velocity, and an analog PD controller pro-vides probe control. A Bruel and Kjaer force sensor
Segment
Bandwidth (Hz)No. Sensors
Fingertip 293 58
Link 2 472 36
Link 1 224 76
Palm 266 64
Table 1: Tactile array type versus bandwidth. between probe and motor measures the applied force, and a Bruel and Kjaer accelerometer measures accel-eration. The mechanical bandwidth of the system is around 50 Hz. Sampling and control is provided by a Micron P90 PC running Labview software, through a ComputerBoards 16-bit ADC.
(Bruel & Kjar)
Linear Motor P-90 LVDT LVVT Accelerometer (Bruel & Kjar)
Force Sensor Digital/Analog I/O
(Entran Devices)
Signal
Sensor Array Generator
(SARCOS)
Figure 6: Experimental setup.
3.2 Static Properties
We experimented on one sensor unit to characterize the static linearity and stiness. Figure 7 shows the hysteresis loop for moderate forces of up to 5 N, which is the expected operating range of the sensor. The amount of hysteresis is small, although the force ver-sus output voltage relation is nonlinear. A segmented straight line t is employed to model this nonlinearity with good results. If substantially higher forces are applied, the hysteresis curves separate signicantly. On the other hand, by investigating the relationship between input force and compression, it is found the
stiness of the sensor unit is around 2:510
4Nm?1
in its operating range.
3.3 Dynamic Properties
A single tactile sensor unit can be modeled as a second order system:
y = mx + b_x + kx (1) 0 1 2 3 4 5 6 7 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Input Force (N) Response (V)
Figure 7: Hysteresis loop (solid lines) plus segmented straight-line curve t (dashed line).
where y and x are sensor response and input force re-spectively and stiness k has already been identied. In our experiment, we kept the probe and the sensor unit well contacted and used load cell readings as in-put. The mass m is the total mass of the sensor unit plus the mass of the probe and half the load cell, which can be precisely measured. Using the Matlab System Identication Toolbox, it is found the system damping is 1:310
2Nsm?1 while the sensor mass is no larger
than 0.05g. This is an over damped system and the bandwidth is no less than 220 Hz. The swept sine test resulted in a at response which indicated the sensor bandwidth is well beyound the 50 Hz bandwidth of the actuator.
The sensor response uctuation and noise level are around 1 mV. For an output range of around 1 V, therefore the dynamic range of the sensor is around 10 bits.
3.4 Spatial Properties
Figure 8 shows the response of a single tactel as the probe is scanned across the surface. The tactel re-sponse changes with probe position due to underlying continuum mechanics of the rubber layers.
One of the most signicant applications of tactile sensing is object localization. We pinpoint more nely the location of a probe by weighted averaging of the responses of neighboring tactels based on the following relation: y = P fixi P fi (2) where y is the location of the probe while xi and fi
respec--2 0 2 4 6 8 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Position of the tip (mm)
Response of sensor unit #15
Figure 8: Single tactel response to lateral scan. tively. Note in our case we only consider one dimen-sion which is sucient to demonstrate the issue.
The results are shown in Figure 9, which compares the location predicted from (2) (dashed line) to the actual probe position set by the x;y stage (solid line). The localization resolution of the palm sensor array is about 1 mm, almost a factor of 3 greater than the tactel spacing. -2 0 2 4 6 8 10 -2 0 2 4 6 8 10 Actual Position (mm) Calculated Position (mm)
Figure 9: Accuracy of point source localization us-ing weighted averages. Solid line: actual probe posi-tion set by x;y stage. Dashed line: predicted posiposi-tion based on weighted average (2)
. 0 0.02 0.04 0.06 0.08 0.1 -5 0 5 10 15 20 25 30 Time (s) Force (N)
Figure 10: Transient force control with loadcell feed-back.
4 Use in Contact Force Control
Because of its dynamic range for force, the tactile sensor can be used as a feedback device in transient force control. In this section, we investigate the per-formance of this control strategy experimentally.
In robot manipulation, transient force control has been a challenging issue. Usually the force controller uses the wrist force/torque sensor as the feedback de-vice. Due to the high freqency disturbances caused by the compliance of the sensor and the manipula-tor itself, the force controller is usually unstable. One of the simplest and most commmonly used methods employs a dominant pole to the system plus lowpass ltering to the force feedback signal [12]. The disad-vantage of this method is sacricing the bandwidth of the controller.
Using a tactile sensor as the force feedback device has two advantages which can result in a much more stable force controller. First, the mass from the sensor to the point of contact is very small and can be ne-glected. Second, the high frequency structure vibra-tion may not be reected in the tactile sensor read-ing due to the high dampread-ing ratio of the sensor. On the other hand, the sensor nonlinearities such as sat-uration and hysterisis become signicant when large transient forces occur, thereby jeopardizing the per-formance of the controller. feedback is not stable.
In our experiment, the actuator command is pre-ltered by a rst order lowpass digital lter with 100 Hz cuto frequency while no lowpass lter is applied to either loadcell or tactile sensor. The force set point is 3.0 N and the threshold for position-force
switch-0 0.02 0.04 0.06 0.08 0.1 -5 0 5 10 15 20 25 Time (s) Force (N)
Figure 11: Transient force control with tactile sensor feedback.
ing is 0.1 N. Figures 10 and 11 show the results of PD control using loadcell and tactile sensor readings as feedbacks, respectively. The controller using tactile sensor feedback works well, while the controller using the loadcell feedback is not stable.
Furthermore we tested the PD controller employing both of the sensors. The controller simply switched to the loadcell from the tactile sensor when the reading from the tactile sensor was beyong a threshold and switched back as soon as the reading was below it. In this case, the threshold was 5.0 N. Figure 12 shows that this strategy results in a better controller.
0 0.02 0.04 0.06 0.08 0.1 -5 0 5 10 15 20 25 30 Time (s) Force (N)
Figure 12: Transient force control with both sensors feedback.
5 Discussion
We have presented a full tactile sensing system for the Utah/MIT Dextrous Hand, which covers nger segments and palm. However, the sensors can be considered as modular systems that can readily be adapted or recast for other end eectors. The sen-sors are constructed of rubber and employ capacitance sensing using oating electrodes, to achieve goals of robustness, manufacturability, and low cost. The tac-tel spacing is 2.77 mm, in order to have large enough electrodes for adequate readings. Local electronics are required for good signal-to-noise ratio.
Experimental results were performed on a at palm array to assess static, dynamic, and spatial properties. The position-force curve is nonlinear and shows mild hysteresis in the operating range. This curve can be well approximated by concatenated straight-line seg-ments. The noise level is 1 mV in a 1 V output range, thus the dynamic range is around 10 bits. The sensor response is at to 50 Hz, the limit of the experimen-tal apparatus; response of up to 1 kHz is expected. Localization of a point probe is improved by weighted averaging of neighboring tactile responses, yielding a factor-of-three improvement to 1 mm.
Preliminary results were presented on the use of the tactile sensor in contact force control. It was shown that contact force is better controlled than through use of a less-local force sensor. However, the combination of tactile plus force sensor yielded the best response, as the dynamic range could be divided between them. These results further the argument of the utility of tactile sensors.
Acknowledgements
The sensor development was supported under the Phase II SBIR Contract NAS 9-18659 from the NASA Johnson Space Center. This research was also partly supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) Network Cen-ters of Excellence (NCE) Institute for Robotics and Intelligent Systems (IRIS) Project MS-1.
References
[1] Boie, R.A., \Capacitive impedance readout tac-tile image sensor," in Proc. 1st IEEE Computer Society Int. Conf. on Robotics, Atlanta, pp. 370-379, March 13-15, 1984.
[2] Dario, P., \Tactile sensing for robots: present and future,"The Robotics Review 1, edited by O. Khatib, J.J. Craig, and T. Lozano-Perez. Cam-bridge, MA: MIT Press, pp. 133-146, 1989. [3] Fearing, R.S., \Tactile sensing mechanisms,"Int.
[4] Grupen, R.A., Henderson, T.C., and McCam-mon, I.D., \A survey of general purpose manip-ulation,"Int. J. Robotics Research, vol. 8, no. 1, pp. 38-62, 1989.
[5] Howe, R.D., and Cutkosky, M.R., \Touch sensing for robotic manipulation and recognition," The Robotics Review 2, edited by O. Khatib, J.J. Craig, and T. Lozano-Perez. Cambridge, MA: MIT Press, pp. 55-112, 1992.
[6] Jacobsen, S.C., Iversen, E.K., Knutti, D.F., Johnson, R.T., and Biggers, K.B., \Design of the
Utah/MIT Dextrous Hand," in Proc. IEEE Int.
Conf. Robotics and Automation, San Francisco, pp. 1520-1532, April 7-10, 1986.
[7] Jacobsen, S.C., McCammon, I.D., Biggers, K.B., and Phillips, R.P., \Design of tactile sensing sys-tems for dextrous manipulators,"IEEE Control Systems Magazine, vol. 8, no. 1, pp. 3-13, 1988. [8] McCammon, I.D., and Jacobsen, S.C., \Tac-tile sensing and control for the Utah/MIT hand," Dextrous Robot Hands, edited by S.T. Venkatara-man and T. Iberall. NY: Springer-Verlag, pp. 239-266, 1990.
[9] Nicholls, H.R., and Lee, M.H., \A survey of robot tactile sensing technology,"Intl. J. Robotics Re-search, vol. 8, no. 3, pp. 3-30, 1989.
[10] Siegel, D.M., Drucker, S.M., and Garabieta, I., \Performance analysis of a tactile sensor," in Proc. IEEE Int. Conf. Robotics and Automation, Raleigh, NC, pp. 1493-1499, March 31-April 3, 1987.
[11] Siegel, D.M., Garabieta, I., and Hollerbach, J.M., \An integrated tactile and thermal sensor," in Proc. IEEE Int. Conf. Robotics and Automation, San Francisco, pp. 1286-1291, April 7-10, 1986. [12] Xu, Y., Hollerbach, J.M., and Ma, D., \A nonlin-ear PD controller for force and contact transient control," IEEE Control Systems, vol. 15, no. 1, pp. 15-21, 1995.
